tools / adapt_config.py
patrickvonplaten's picture
up
62fc896
raw
history blame
No virus
8.98 kB
import argparse
import json
import os
import shutil
from tempfile import TemporaryDirectory
from typing import List, Optional
from huggingface_hub import CommitInfo, CommitOperationAdd, Discussion, HfApi, hf_hub_download
from huggingface_hub.file_download import repo_folder_name
class AlreadyExists(Exception):
pass
def is_index_stable_diffusion_like(config_dict):
if "_class_name" not in config_dict:
return False
compatible_classes = [
"AltDiffusionImg2ImgPipeline",
"AltDiffusionPipeline",
"CycleDiffusionPipeline",
"StableDiffusionImageVariationPipeline",
"StableDiffusionImg2ImgPipeline",
"StableDiffusionInpaintPipeline",
"StableDiffusionInpaintPipelineLegacy",
"StableDiffusionPipeline",
"StableDiffusionPipelineSafe",
"StableDiffusionUpscalePipeline",
"VersatileDiffusionDualGuidedPipeline",
"VersatileDiffusionImageVariationPipeline",
"VersatileDiffusionPipeline",
"VersatileDiffusionTextToImagePipeline",
"OnnxStableDiffusionImg2ImgPipeline",
"OnnxStableDiffusionInpaintPipeline",
"OnnxStableDiffusionInpaintPipelineLegacy",
"OnnxStableDiffusionPipeline",
"StableDiffusionOnnxPipeline",
"FlaxStableDiffusionPipeline",
]
return config_dict["_class_name"] in compatible_classes
def convert_single(model_id: str, folder: str) -> List["CommitOperationAdd"]:
config_file = "scheduler/scheduler_config.json"
os.makedirs(os.path.join(folder, "scheduler"), exist_ok=True)
model_index_file = hf_hub_download(repo_id=model_id, filename="model_index.json")
with open(model_index_file, "r") as f:
index_dict = json.load(f)
if not is_index_stable_diffusion_like(index_dict):
print(f"{model_id} is not of type stable diffusion.")
return False, False
old_config_file = hf_hub_download(repo_id=model_id, filename=config_file)
new_config_file = os.path.join(folder, config_file)
success = convert_file(old_config_file, new_config_file)
if success:
operations = [CommitOperationAdd(path_in_repo=config_file, path_or_fileobj=new_config_file)]
model_type = success
return operations, model_type
else:
return False, False
def convert_file(
old_config: str,
new_config: str,
):
with open(old_config, "r") as f:
old_dict = json.load(f)
if "clip_sample" not in old_dict:
print("Make scheduler DDIM compatible")
old_dict["clip_sample"] = False
else:
print("No matching config")
return False
# is_stable_diffusion = "down_block_types" in old_dict and list(old_dict["down_block_types"]) == ["CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D"]
#
# is_stable_diffusion_1 = is_stable_diffusion and ("use_linear_projection" not in old_dict or old_dict["use_linear_projection"] is False)
# is_stable_diffusion_2 = is_stable_diffusion and ("use_linear_projection" in old_dict and old_dict["use_linear_projection"] is True)
#
# if not is_stable_diffusion_1 and not is_stable_diffusion_2:
# print("No matching config")
# return False
#
# if is_stable_diffusion_1:
# if old_dict["sample_size"] == 64:
# print("Dict correct")
# return False
#
# print("Correct stable diffusion 1")
# old_dict["sample_size"] = 64
#
# if is_stable_diffusion_2:
# if old_dict["sample_size"] == 96:
# print("Dict correct")
# return False
#
# print("Correct stable diffusion 2")
# old_dict["sample_size"] = 96
#
with open(new_config, 'w') as f:
json_str = json.dumps(old_dict, indent=2, sort_keys=True) + "\n"
f.write(json_str)
#
# return "Stable Diffusion 1" if is_stable_diffusion_1 else "Stable Diffusion 2"
return "Stable Diffusion"
def previous_pr(api: "HfApi", model_id: str, pr_title: str) -> Optional["Discussion"]:
try:
discussions = api.get_repo_discussions(repo_id=model_id)
except Exception:
return None
for discussion in discussions:
if discussion.status == "open" and discussion.is_pull_request and discussion.title == pr_title:
return discussion
def convert(api: "HfApi", model_id: str, force: bool = False) -> Optional["CommitInfo"]:
# pr_title = "Correct `sample_size` of {}'s unet to have correct width and height default"
pr_title = "Add `clip_sample=False` to scheduler to make model compatible with DDIM."
info = api.model_info(model_id)
filenames = set(s.rfilename for s in info.siblings)
if "unet/config.json" not in filenames:
print(f"Model: {model_id} has no 'unet/config.json' file to change")
return
if "vae/config.json" not in filenames:
print(f"Model: {model_id} has no 'vae/config.json' file to change")
return
with TemporaryDirectory() as d:
folder = os.path.join(d, repo_folder_name(repo_id=model_id, repo_type="models"))
os.makedirs(folder)
new_pr = None
try:
operations = None
pr = previous_pr(api, model_id, pr_title)
if pr is not None and not force:
url = f"https://huggingface.co/{model_id}/discussions/{pr.num}"
new_pr = pr
raise AlreadyExists(f"Model {model_id} already has an open PR check out {url}")
else:
operations, model_type = convert_single(model_id, folder)
if operations:
pr_title = pr_title.format(model_type)
# if model_type == "Stable Diffusion 1":
# sample_size = 64
# image_size = 512
# elif model_type == "Stable Diffusion 2":
# sample_size = 96
# image_size = 768
# pr_description = (
# f"Since `diffusers==0.9.0` the width and height is automatically inferred from the `sample_size` attribute of your unet's config. It seems like your diffusion model has the same architecture as {model_type} which means that when using this model, by default an image size of {image_size}x{image_size} should be generated. This in turn means the unet's sample size should be **{sample_size}**. \n\n In order to suppress to update your configuration on the fly and to suppress the deprecation warning added in this PR: https://github.com/huggingface/diffusers/pull/1406/files#r1035703505 it is strongly recommended to merge this PR."
# )
contributor = model_id.split("/")[0]
pr_description = (
f"Hey {contributor} 👋, \n\n Your model repository seems to contain a stable diffusion checkpoint. We have noticed that your scheduler config currently does not correctly work with the [DDIMScheduler](https://huggingface.co/docs/diffusers/main/en/api/schedulers#diffusers.DDIMScheduler) because `clip_sample` is not set to False and will therefore [incorrectly default to True](https://github.com/huggingface/diffusers/blob/3ce6380d3a2ec5c3e3f4f48889d380d657b151bc/src/diffusers/schedulers/scheduling_ddim.py#L127). \n The official stable diffusion checkpoints have `clip_sample=False` so that the scheduler config works will **all** schedulers, see: https://huggingface.co/stabilityai/stable-diffusion-2-1-base/blob/main/scheduler/scheduler_config.json#L7. \n\n We strongly recommend that you merge this PR to make sure your model works correctly with DDIM. \n\n Diffusingly, \n Patrick."
)
new_pr = api.create_commit(
repo_id=model_id,
operations=operations,
commit_message=pr_title,
commit_description=pr_description,
create_pr=True,
)
print(f"Pr created at {new_pr.pr_url}")
else:
print(f"No files to convert for {model_id}")
finally:
shutil.rmtree(folder)
return new_pr
if __name__ == "__main__":
DESCRIPTION = """
Simple utility tool to convert automatically some weights on the hub to `safetensors` format.
It is PyTorch exclusive for now.
It works by downloading the weights (PT), converting them locally, and uploading them back
as a PR on the hub.
"""
parser = argparse.ArgumentParser(description=DESCRIPTION)
parser.add_argument(
"model_id",
type=str,
help="The name of the model on the hub to convert. E.g. `gpt2` or `facebook/wav2vec2-base-960h`",
)
parser.add_argument(
"--force",
action="store_true",
help="Create the PR even if it already exists of if the model was already converted.",
)
args = parser.parse_args()
model_id = args.model_id
api = HfApi()
convert(api, model_id, force=args.force)